Telco Assessment

Problem Assessment

Our assessment consists of on a geolocation information of a set of 1,752 wireless subscribers for a phone company in various cells across Nova Scotia. The historical data was provided as of September 1st, 2020. We have also been provided a dataset containing information for the Poseidon CTC Mall and its geofence cell list.

By using the data science toolbox, we need to understand how much average time the subscribers are spending at the Poseidon CTC Mall. The goal of the project is to create an executive summary and create business insights that will help the marketing team to achieve this objective.

Business Problem

Recently, there has been several discussions within our marketing team to develop a strategy to approaching our clients more efficiently, accordingly to the time that they spend in the current Poseidon CTC Mall. As a team, we need to understand what should be the better time frame in which we can target those clients by using the average dwell time, which is a metric to show the “amount of time” a device uses a particular channel.

Our current point of interest (POI) to develop this strategy, corresponds to the Poseidon CTC Mall.

Business Value

By developing a strategy to target our clients efficiently, we are creating organic growth to the business which is achieved by enhancing our sales revenue. By putting in production our final data product, we will be able to get to know our clients better and improve our Voice of Customer.

Tools used

The tools used for this project are:

Python Programming: Python provides a vast set of utilities do develop data science and machine learning through the most prominent packages: Pandas, Scikit-Learn, Numpy and others.

R - Reticulate: Reticulate is a tool to run both Python and R on the same environment

R Markdown: R Programming (through RStudio) provides a set of tools to publish our work using documents and HTML sources, to provide better presentations to our executives or audience.

ASSESSMENT PART 1

1. Loading packages

First we load reticulate package to write Python / R code in our Markdown enviroment:

# Use only once:

#Sys.setenv(RETICULATE_PYTHON = "/usr/local/bin/python3.7")
#library(reticulate)
#use_python('/usr/bin/python3', require = T)

Second, we import Python packages

import pandas as pd              # package for data wrangling and transformation
import numpy as np               # package for linear algebra
import seaborn as sns            # advanced graphical interface package 
import matplotlib.pyplot as plt  # basic graphical interface package
import copy
import warnings
warnings.filterwarnings('ignore')
from sklearn.cluster import KMeans

2. Collecting & Preparing Data

The fist step is to get the data which has been stored in a csv format. In addition to the files already provided, I was able to find supporting information that would help our analysis at https://mcc-mnc-list.com/list.

With the cgi dataset I was able to get information about:

MCC: Mobile Country Code MNC: Mobile Network Code LAC: Location Area Code CI: Cell Identity

# geo location data 
geo = pd.read_csv('/home/analytics/R/Projects/Python/datasets/telco/geolocation_data.csv')

# Poseidon CTC Mall Geofence data
poseidon = pd.read_csv('/home/analytics/R/Projects/Python/datasets/telco/Poseidon CTC Mall Geofence Cell List.csv')

# cgi data
cgi = pd.read_csv('/home/analytics/R/Projects/Python/datasets/telco/cgi.csv')

print(geo.shape)

## (408917, 3)

geo.head()

##                                              imsi_id  ...                 cgi_id
## 0  778fb33d64b6dcbc67c79ca6d35f7820f993f26c74502c...  ...  4G:302-220-113919-131
## 1  753c0c75ac582307c3b0d04bbe4cfe7d118684d1b4060e...  ...   4G:302-220-113155-41
## 2  05f8982252e683f1fed1f3ebd42f60848587a4e270ddcc...  ...   4G:302-220-113937-10
## 3  d5ce6f7f332850a23ae4700231c87dec634a1e7832304d...  ...    302-220-11201-23458
## 4  05f8982252e683f1fed1f3ebd42f60848587a4e270ddcc...  ...   4G:302-220-113937-10
## 
## [5 rows x 3 columns]

print(poseidon.shape)

## (105, 1)

poseidon.head()

##   Poseidon CTC Mall Geofence Cell List
## 0                4G:302-220-113122-132
## 1                 4G:302-220-113122-42
## 2                4G:302-220-113122-112
## 3                  4G:302-220-113122-2
## 4                4G:302-220-113122-131

poseidon['Poseidon CTC Mall Geofence Cell List'].nunique()

## 75

poseidon.drop_duplicates(subset=['Poseidon CTC Mall Geofence Cell List'], inplace = True)

poseidon.shape # indeed we have 75 cells

## (75, 1)

poseidon.head() 

##   Poseidon CTC Mall Geofence Cell List
## 0                4G:302-220-113122-132
## 1                 4G:302-220-113122-42
## 2                4G:302-220-113122-112
## 3                  4G:302-220-113122-2
## 4                4G:302-220-113122-131

cgi.shape

## (3, 6)

cgi.head()

##    MCC  ...                         Operator
## 0  302  ...                   Telus Mobility
## 1  302  ...                    Bell Mobility
## 2  302  ...  Shared Telus, Bell, and SaskTel
## 
## [3 rows x 6 columns]

cgi.MCC = cgi.MCC.astype('object')
cgi.MNC = cgi.MNC.astype('object')

3. Merging the data

df = pd.merge(poseidon, geo, left_on = 'Poseidon CTC Mall Geofence Cell List', right_on = 'cgi_id')

df.head()

##   Poseidon CTC Mall Geofence Cell List  ...                 cgi_id
## 0                4G:302-220-113122-132  ...  4G:302-220-113122-132
## 1                4G:302-220-113122-132  ...  4G:302-220-113122-132
## 2                4G:302-220-113122-132  ...  4G:302-220-113122-132
## 3                4G:302-220-113122-132  ...  4G:302-220-113122-132
## 4                 4G:302-220-113122-42  ...   4G:302-220-113122-42
## 
## [5 rows x 4 columns]

df.shape

## (432, 4)

No of unique clients to be analized: 33

df.imsi_id.nunique()

## 33

4. Data verification / cleaning

telco_nulls = df.isna().sum()[df.isna().sum()!= 0].reset_index().rename(columns=
                                                                        {'index': 'col_name', 0: '#'})
telco_nulls['%'] = telco_nulls['#'] / df.shape[0]
telco_nulls

## Empty DataFrame
## Columns: [col_name, #, %]
## Index: []

No NULL values were found on the data.

5. Data Transformation & Feature Engineering

new = df["cgi_id"].str.split("-", n = 4, expand = True)
broad_MCC = pd.DataFrame(new[0])
# If broadband does not contain 4G then place "other"
broad_MCC[1] = broad_MCC[0].map(lambda x: "4G" if "4G" in x else 'Other')

broad_MCC[2] = broad_MCC[0].str[-3:]
broad_MCC[2].nunique() #indeed we have only 1 value.

## 1

broad_MCC.head()

##         0   1    2
## 0  4G:302  4G  302
## 1  4G:302  4G  302
## 2  4G:302  4G  302
## 3  4G:302  4G  302
## 4  4G:302  4G  302

df['broadband'] = broad_MCC[1]
df['MCC'] = broad_MCC[2]
df['MNC'] =  new[1]
df['LAC'] =  new[2]
df['CI'] =  new[3]
df.head()

##   Poseidon CTC Mall Geofence Cell List  ...   CI
## 0                4G:302-220-113122-132  ...  132
## 1                4G:302-220-113122-132  ...  132
## 2                4G:302-220-113122-132  ...  132
## 3                4G:302-220-113122-132  ...  132
## 4                 4G:302-220-113122-42  ...   42
## 
## [5 rows x 9 columns]

df.drop(columns=['Poseidon CTC Mall Geofence Cell List', 'cgi_id'], inplace = True)
df.head()

##                                              imsi_id  ...   CI
## 0  b58217bd1fe520ae4cb6a2228ea1cffa68590965cc5ba8...  ...  132
## 1  2f65da3d1f0ed00d831a2037868732a262523d6231e9bf...  ...  132
## 2  450c1b07f639674e3e004251fb86ce273ac21bc1c1754e...  ...  132
## 3  2f627a032a70455a80760c51c0c8cc9282c87f52a25703...  ...  132
## 4  044aa99b8a013c64e3924cf76dec7c9a25d7315bcaccaf...  ...   42
## 
## [5 rows x 7 columns]

df.MCC = df.MCC.astype('str')
df.MNC = df.MNC.astype('str')
cgi.MCC = cgi.MCC.astype('str')
cgi.MNC = cgi.MNC.astype('str')
telco = pd.merge(df, cgi, how = 'left', left_on = ['MCC', 'MNC'], right_on = ['MCC', 'MNC' ])
telco.head()

##                                              imsi_id  ...        Operator
## 0  b58217bd1fe520ae4cb6a2228ea1cffa68590965cc5ba8...  ...  Telus Mobility
## 1  2f65da3d1f0ed00d831a2037868732a262523d6231e9bf...  ...  Telus Mobility
## 2  450c1b07f639674e3e004251fb86ce273ac21bc1c1754e...  ...  Telus Mobility
## 3  2f627a032a70455a80760c51c0c8cc9282c87f52a25703...  ...  Telus Mobility
## 4  044aa99b8a013c64e3924cf76dec7c9a25d7315bcaccaf...  ...  Telus Mobility
## 
## [5 rows x 11 columns]

telco.shape

## (432, 11)

telco.drop(columns=['ISO', 'Country', 'Brand'], inplace = True)

telco.head()

##                                              imsi_id  ...        Operator
## 0  b58217bd1fe520ae4cb6a2228ea1cffa68590965cc5ba8...  ...  Telus Mobility
## 1  2f65da3d1f0ed00d831a2037868732a262523d6231e9bf...  ...  Telus Mobility
## 2  450c1b07f639674e3e004251fb86ce273ac21bc1c1754e...  ...  Telus Mobility
## 3  2f627a032a70455a80760c51c0c8cc9282c87f52a25703...  ...  Telus Mobility
## 4  044aa99b8a013c64e3924cf76dec7c9a25d7315bcaccaf...  ...  Telus Mobility
## 
## [5 rows x 8 columns]

6. Exploratory Data Analysis

Our data exploration consists in analyzing our categorical & numerical variables. Next, we will be focusing on the time-based analysis.

telco.imsi_id.nunique()

## 33

telco.shape[0]

## 432

sns.countplot(telco.broadband)

round(100 * telco.groupby(['broadband']).size() / telco.shape[0])

## broadband
## 4G       34.0
## Other    66.0
## dtype: float64

telco.groupby(['MCC']).size()

## MCC
## 302    432
## dtype: int64

telco.groupby(['MNC']).size()

## MNC
## 220    431
## 880      1
## dtype: int64

sns.countplot(telco.LAC)
plt.show()

LAC = pd.DataFrame()
LAC['n'] = telco.groupby(['LAC']).size()
LAC['%'] = round(100 * (LAC['n'] / telco.shape[0]))
LAC

##           n     %
## LAC              
## 11204   285  66.0
## 113122  104  24.0
## 123353   43  10.0

CI = pd.DataFrame()
CI['n'] = telco.groupby(['CI']).size().sort_values(ascending = False)
CI['%'] = round(100 * (CI['n'] / telco.shape[0]))
CI

##          n     %
## CI              
## 24122  111  26.0
## 24353   79  18.0
## 21122   50  12.0
## 41      48  11.0
## 25122   43  10.0
## 111     37   9.0
## 131     23   5.0
## 1       18   4.0
## 42       5   1.0
## 132      4   1.0
## 2        4   1.0
## 112      3   1.0
## 71       3   1.0
## 22122    2   0.0
## 11       1   0.0
## 22       1   0.0

telco.event_ts = df.event_ts.astype('datetime64[ns]')
telco = telco.sort_values(by = ['event_ts'], ascending = True)
time = pd.DatetimeIndex(telco.event_ts)  # converts to datetime object index
telco['hours'] = (time.hour  + time.minute / 60 + time.second / 3600 )
telcov2 = copy.deepcopy(telco)
telco.head(10)

##                                                imsi_id  ...     hours
## 61   9f6dbb4957e34ad65b4061e1ac326afa539752db6a1f57...  ...  0.000000
## 52   9f6dbb4957e34ad65b4061e1ac326afa539752db6a1f57...  ...  0.000000
## 70   9f6dbb4957e34ad65b4061e1ac326afa539752db6a1f57...  ...  0.000000
## 149  9dd48ceccaea12ad1ed02afa45f096bba2409f1f5e04d5...  ...  0.550000
## 148  9dd48ceccaea12ad1ed02afa45f096bba2409f1f5e04d5...  ...  0.550000
## 165  9dd48ceccaea12ad1ed02afa45f096bba2409f1f5e04d5...  ...  0.566667
## 166  9dd48ceccaea12ad1ed02afa45f096bba2409f1f5e04d5...  ...  0.566667
## 171  e41a1559c092dfa12acf39072821ce575665fdc4c2a46c...  ...  0.616667
## 186  d6a5b12c28dc324c1ff61b2540336c0fbe3ac89b1744be...  ...  0.800000
## 155  d6a5b12c28dc324c1ff61b2540336c0fbe3ac89b1744be...  ...  0.800000
## 
## [10 rows x 9 columns]

output = telco.groupby('imsi_id').agg(
    dwell_time = ('hours', 'mean' ),
    r_visits = ('hours', 'count')
     ).sort_values(by = ['dwell_time'], ascending = True).reset_index()

pd.set_option('display.max_rows', 50)
output

##                                               imsi_id  dwell_time  r_visits
## 0   9dd48ceccaea12ad1ed02afa45f096bba2409f1f5e04d5...    0.558333         4
## 1   d6a5b12c28dc324c1ff61b2540336c0fbe3ac89b1744be...    0.800000         2
## 2   9366f7e222a3682413fc4b3480289bf5ee8bee9f5c9354...    1.316667         1
## 3   f48b2e8df1f9f912c788b57e6f2c3246e08b74440f033e...    1.966667         6
## 4   f70cd60e8a80eebd2422c0db3d8fb5d9b441bb6ef22c71...   10.123333         5
## 5   e41a1559c092dfa12acf39072821ce575665fdc4c2a46c...   10.822917        16
## 6   10ef0c74bc117442d7af58334e1a7f6d7e6e374a705e1f...   11.524242        11
## 7   9f6dbb4957e34ad65b4061e1ac326afa539752db6a1f57...   12.148148         9
## 8   c6f43f7ae49f220a883f6b63b26024eff38ce3c2e773a6...   13.783965       132
## 9   e2ee0f057e15ceedbb10b6c7ba335487fd20cfb7b1ca55...   13.992857         7
## 10  12f577a9e7c23830cfc662a8954126c96a2f966c248457...   14.255556        12
## 11  d993bd44520ea984f45182b7398c98a2af4e0da09ef423...   14.733333         1
## 12  321b7cf31816a63fb196a208c3eeea20de04cd59dc35d8...   15.870000         5
## 13  044aa99b8a013c64e3924cf76dec7c9a25d7315bcaccaf...   15.920833         4
## 14  cbc12ecb1aa2a955a6426a3114f1a3230ff618560be98a...   16.233333         5
## 15  2112b7c11b6fd3e8b5c78af39566e4834685bdcaf56c6b...   16.600000         4
## 16  cfef7cf475779a432e73e64641bc5d2abb58e5eebc15ce...   17.488889         3
## 17  7948cc9c2c4ede47301ef73252a41fdf9b8e07817f5604...   17.503604        37
## 18  2f65da3d1f0ed00d831a2037868732a262523d6231e9bf...   17.633333         2
## 19  16d323326378966732abaafddad4027288389ee110cbd1...   17.683333         3
## 20  f85657aee6b4977e8909ea8de56f26d3d7010a40e5d210...   17.773232        33
## 21  f043c2b9f621827a588b37bf25f7a28f41ce3ce5928dbb...   18.033333         1
## 22  b58217bd1fe520ae4cb6a2228ea1cffa68590965cc5ba8...   18.561538        13
## 23  05b5aa3eb6db1abda56d8cd0f15e6f5ce2307d16cc9e05...   19.327778         3
## 24  9b03a42cf240abbb766ddbe3e5091bfc1cc33c0bf08409...   19.576316        19
## 25  2f627a032a70455a80760c51c0c8cc9282c87f52a25703...   19.905556        18
## 26  2acb2650d37de9efe5ef71d02c0f9ac8dba554f6f70370...   19.916667        12
## 27  16a2a5883d4b3b6f5d4f688e43887bbf712769c612b9da...   19.983333         1
## 28  1581807a0d2e3c44b53791d8601ae83e0e1ac51142d84d...   21.066667         2
## 29  bc9e5df525777dbe77e262794de72923365aa37b3c019f...   21.129902        34
## 30  85d444651f5123ca9e97d30ff30a40f0a65d0487ac37df...   21.250000         4
## 31  450c1b07f639674e3e004251fb86ce273ac21bc1c1754e...   23.471111        15
## 32  059e71d6579dadcfa0cef78ee779b671ac254549a2b4bb...   23.843750         8

output[output.r_visits == output.r_visits.max()]

##                                              imsi_id  dwell_time  r_visits
## 8  c6f43f7ae49f220a883f6b63b26024eff38ce3c2e773a6...   13.783965       132

output.nlargest(3, 'r_visits')

##                                               imsi_id  dwell_time  r_visits
## 8   c6f43f7ae49f220a883f6b63b26024eff38ce3c2e773a6...   13.783965       132
## 17  7948cc9c2c4ede47301ef73252a41fdf9b8e07817f5604...   17.503604        37
## 29  bc9e5df525777dbe77e262794de72923365aa37b3c019f...   21.129902        34

7. DWELL TIME AVERAGE ANALYSYS

output_bin = copy.deepcopy(output)

output_bin['points_bin'] = pd.qcut(output_bin['dwell_time'], q=8)

#bining the data
output_bin = output_bin.groupby('points_bin').agg(
    r_visits = ('r_visits', 'sum')
     ).reset_index()

#plotting the data:     
plt.figure(figsize = (15, 5))
sns.barplot(x = output_bin.points_bin, y = output_bin.r_visits, color = 'blue')
# Labeling of the plot
plt.xlabel('dwell time (hrs)') 
plt.ylabel('Recurrent visits')
plt.title('Distribution of dwell time (Poseidon CTC Mall)')
plt.show()

plt.figure(figsize = (15, 5))
sns.scatterplot(x = output.dwell_time, y = output.r_visits)
# Labeling of the plot
plt.xlabel('dwell time (hrs)') 
plt.ylabel('Recurrent visits')
plt.title('Dwell_time vs Recurrent visits (Poseidon CTC Mall)')
plt.show()

output[output.r_visits == output.r_visits.max()]

##                                              imsi_id  dwell_time  r_visits
## 8  c6f43f7ae49f220a883f6b63b26024eff38ce3c2e773a6...   13.783965       132

ts = copy.deepcopy(telcov2)
ts['cnt'] = 1
#ts = ts.groupby('event_ts').agg({"cnt": "sum"}).reset_index()
ts.event_ts = ts.event_ts.astype(str)

8. Client Trend Analysis

library(reticulate)
ts_data <- py$ts

# Forecasting Libraries ----
library(forecast)    
library(tidymodels)   
library(modeltime)    
library(tidyverse)    
library(lubridate)    
library(timetk)       
library(plotly)

#renaming colums

ts_data$Date <- ts_data$event_ts 

#converting to datetime
ts_data$Date <- as.POSIXct(as.character(ts_data$Date), format = "%Y-%d-%m %H:%M:%S")
ts_data <- as_tibble(ts_data)

#summarize by hour

p1 <- ts_data %>%  group_by(Date) %>%
  summarize_by_time(.date_var=Date, .by = 'hour', no_visits = sum(cnt) ) %>%
  plot_time_series(Date, no_visits,  .smooth = FALSE,
                   .title = "Poseidon CTC Malls Visits Trend - Sep 1st 2020", .y_lab =  "count of client visits" ) 
p1

#summarize by hour

p2 <- ts_data %>%  group_by(broadband) %>%
  summarize_by_time(.date_var=Date, .by = 'hour', no_visits = sum(cnt) ) %>%
  plot_time_series(Date, no_visits,  .smooth = TRUE,
                   .color_var = broadband ,
                   .title = "Poseidon CTC Broadband types - Sep 1st 2020", 
                   .y_lab =  "count of client visits", .facet_ncol = 2,
                   .facet_scales = "free" )  
p2

#summarize by hour

p3 <- ts_data %>%  group_by(LAC) %>%
  summarize_by_time(.date_var=Date, .by = 'hour', no_visits = sum(cnt) ) %>%
  plot_time_series(Date, no_visits,  
                   .title = "Poseidon CTC Mall by LAC - Sep 1st 2020", 
                   .y_lab =  "count of client visits", .facet_ncol = 2,
                   .facet_scales = "free" ) 
p3

#summarize by hour

p4 <- ts_data %>%  group_by(CI) %>%
  summarize_by_time(.date_var=Date, .by = 'hour', no_visits = sum(cnt) ) %>% arrange(desc(no_visits)) %>% 
  filter(CI %in% c('111', '24122', '131', '24122', '41')) %>%
  plot_time_series(Date, no_visits,  .smooth = FALSE,
                   .title = "Poseidon CTC Mall by CI - Sep 1st 2020", 
                   .y_lab =  "count of client visits", .facet_ncol = 1,
                   .facet_scales = "free" ) 
p4

ASSESSMENT PART 2

Once the average dwell times analysis is complete, cluster the subscribers seen inside Poseidon CTC Mall into various segments, and explain the various segments. We would also like to know how did you select the #segments (or K value), if at all.

1. Selecting the data

Storing the data for the columns: dwell_time and r_visits (by each distinct client):

output_2 = output
X = output_2.iloc[:, [1,2]].values  

2. Model Building

# By applying K-Means algorithm, we are going to find the best cluster number 
# The idea, is to reduce WCSS (Within - Cluster - Sum of Squares)
from sklearn.cluster import KMeans
wcss = []

# Getting the maximun numbers of clusters (to begin with, we set 10 by default)

# According to the official documentation of scikit learn, we use:

#‘k-means++’ : selects initial cluster centers for k-mean clustering in a smart way to speed up 
#convergence.

# random_state = 123, just be sure a fixed randomness

# Inertia tell us calculates the sum of distances of all the points within a cluster from the centroid of that
# cluster

# by default, we chose to select an 'auto' algorithm for converging. Euclidean distance is used in this case,
# but it might change according to our performance.

for i in range(1,11):
    kmeans = KMeans(n_clusters= i, max_iter = 300, 
                    init='k-means++', random_state=123, 
                    algorithm='auto')
    kmeans.fit(X)
    wcss.append(kmeans.inertia_)

## KMeans(n_clusters=1, random_state=123)
## KMeans(n_clusters=2, random_state=123)
## KMeans(n_clusters=3, random_state=123)
## KMeans(n_clusters=4, random_state=123)
## KMeans(n_clusters=5, random_state=123)
## KMeans(n_clusters=6, random_state=123)
## KMeans(n_clusters=7, random_state=123)
## KMeans(random_state=123)
## KMeans(n_clusters=9, random_state=123)
## KMeans(n_clusters=10, random_state=123)

3. Visualizing results

# Viewing ELBOW method to get the optimal value of K

plt.plot(range(1,11), wcss)
plt.title('Elbow Method')
plt.xlabel('No. of clusters')
plt.ylabel('wcss')
plt.show() 

From the graph above, we can see that according to the “Elbow validation approach”, if we choose 6 clusters, then we will achieving a minium error of wcss

#Model Build
kmeans_model = KMeans(n_clusters= 6, max_iter = 300, 
                    init='k-means++', random_state=123, 
                    verbose = 1, algorithm='auto')
output_2 ['y_pred'] = kmeans_model.fit_predict(X)

## Initialization complete
## Iteration 0, inertia 476.2443094109775
## Iteration 1, inertia 305.4512442082231
## Converged at iteration 1: strict convergence.
## Initialization complete
## Iteration 0, inertia 545.1293060336468
## Iteration 1, inertia 312.2361716096957
## Iteration 2, inertia 306.7580124449414
## Converged at iteration 2: strict convergence.
## Initialization complete
## Iteration 0, inertia 450.11941478483624
## Iteration 1, inertia 371.25146932287197
## Iteration 2, inertia 315.6318486261001
## Iteration 3, inertia 306.7580124449414
## Converged at iteration 3: strict convergence.
## Initialization complete
## Iteration 0, inertia 437.3813717106833
## Iteration 1, inertia 354.67313867714694
## Iteration 2, inertia 306.7580124449414
## Converged at iteration 2: strict convergence.
## Initialization complete
## Iteration 0, inertia 483.9692655293205
## Iteration 1, inertia 360.8337449127447
## Iteration 2, inertia 310.94913777733245
## Iteration 3, inertia 305.45124420822316
## Converged at iteration 3: strict convergence.
## Initialization complete
## Iteration 0, inertia 544.6388108996249
## Iteration 1, inertia 354.6731386771469
## Iteration 2, inertia 306.7580124449414
## Converged at iteration 2: strict convergence.
## Initialization complete
## Iteration 0, inertia 448.99713376178613
## Iteration 1, inertia 306.7580124449414
## Converged at iteration 1: strict convergence.
## Initialization complete
## Iteration 0, inertia 408.68816905742136
## Iteration 1, inertia 306.7580124449414
## Converged at iteration 1: strict convergence.
## Initialization complete
## Iteration 0, inertia 548.0569280821813
## Iteration 1, inertia 404.6820345169024
## Iteration 2, inertia 383.5558342152465
## Iteration 3, inertia 373.7482806253318
## Converged at iteration 3: strict convergence.
## Initialization complete
## Iteration 0, inertia 520.4707242606468
## Iteration 1, inertia 346.1610231213367
## Iteration 2, inertia 336.1378259792391
## Iteration 3, inertia 311.73039202918505
## Iteration 4, inertia 306.7580124449414
## Converged at iteration 4: strict convergence.

# plot results
plt.figure(figsize=(10,6))

plot_clusters = sns.scatterplot(x='dwell_time', y='r_visits', hue='y_pred',palette='Set1', s=100, alpha=0.2,  
                data=output_2).set_title('KMeans Clusters (6)', fontsize=15)


plt.show()

By looking at the previous graph we are able to segment our customers using 4 clusters. It’s important to remark the most of our clients stay below 40 visits a day. There is one particular cluster (1) that was highly skewed greater than 120. This shows that from now on, any other potential client which behaves similarly will belong to this particular cluster.

FINAL WORDS AND RECOMMENDATIONS

(1) By analyzing the performance of our Poseidon CTC PO dwell time, we were able to understand the patterns and relationships on our data that will help the marketing department to know and approach better to our clients. We were capable to segment the most prominent clients in our current portfolio. This will also allow the sales leaders to focus specifically on segments in which they can target new customers or making offerings to the current ones.

(2) By understanding our broadband / MNC / LAC, we were able to find the frequency and cyclical patterns in our client behavior. That would help the marketing leaders to focus specifically in particular times in the current day.

(3) We used machine learning to find a solution in clustering our current clients according to their dwell time and count of visits. If there are any new clients, our algorithm will be capable to assign a cluster to a particular client. It’s highly recommended that we keep up updating our models periodically, since the structure, behavior of the clients and systematic events could affect the performance of those models.

(4) Next steps are to publish put our machine learning in production or embed the results into a dashboard. As per the time series analysis, is also recommended to build a forecasting model capable to predict the number of clients demanded in our current Poseidon CTC PO.